That code linked in 1. looks like it should have been deleted when we changed controller manager to fill in the replicas field. Can we take that out ASAP?

I think having watchers get updated "only" every 10 seconds is not actually a problem at this point in time.

That happened automatically when we started using generic etcd: https://github.com/GoogleCloudPlatform/kubernetes/blob/master/pkg/master/master.go#L380. I forgot to delete the code in registry/etcd, will do.

Ah.

So is a fair TL;DR: "is a 10 second refresh for rc.status.replicas fast enough?"

Yes, I get the feeling it's not because it could be 20s before it is accurate right now, which slows down stop (even though stop will soon use watch, eliminating its own 3s poll interval).

We can probably make it as fast as we want when we stop listing all the things from the apiserver.

Correction: In spite of the 20s lag in the general case stop will only be slowed down by 10s, because resizing the rc to 0 will trigger a sync loop.

To answer my own question: Yes, 10 seconds sounds totally reasonable. I think this is working now, and the parts that aren't perfect are covered by other issues.